Volume 4, Issue 3, Pages 533-543 (March 2018) Using Evolutionary Algorithms and Machine Learning to Explore Sequence Space for the Discovery of Antimicrobial Peptides  Mari Yoshida, Trevor Hinkley, Soichiro Tsuda, Yousef M. Abul-Haija, Roy T. McBurney, Vladislav Kulikov, Jennifer S. Mathieson, Sabrina Galiñanes Reyes, Maria D. Castro, Leroy Cronin  Chem  Volume 4, Issue 3, Pages 533-543 (March 2018) DOI: 10.1016/j.chempr.2018.01.005 Copyright © 2018 Terms and Conditions

Chem 2018 4, 533-543DOI: (10.1016/j.chempr.2018.01.005) Copyright © 2018 Terms and Conditions

Figure 1 Optimization of Antimicrobial Peptides (A) IC50 values of all the peptides in each generation are shown in a boxplot. Each solid line in the boxplot represents the median IC50 value. Boxes show the first and third quartiles. The upper and lower whiskers indicate 50% of the values higher and lower, respectively, than the median. Black dots represent outliers. Generations A and B, generations 1–3, strong substitutions, and control generations 1–3 are shown in green, yellow, blue, and red, respectively. The IC50 value of the WT peptide is shown by a dashed line. (B) Fold changes of IC50 values in comparison with the WT sequence. Barplots indicate average changes in IC50 values with standard errors. The range of activity of the most potent peptides in each generation is shown by an asterisk. (C) A boxplot of the fitness matrix values for each generation (G-B, G1, and G2). The maximum and average values for each generation are shown as blue and red circles, respectively. Chem 2018 4, 533-543DOI: (10.1016/j.chempr.2018.01.005) Copyright © 2018 Terms and Conditions

Figure 2 Changes in the Physicochemical Properties of Peptides Values of the properties of peptides were subtracted from those of the WT peptide. Average changes in (A) net charge and (B) hydrophobic moment in individual generations are shown in barplots with standard errors. Changes in the most potent peptides in individual generations are shown as asterisks. Generations A and B, generations 1–3, and control generation 1–3 are shown in green, yellow, and red, respectively. Chem 2018 4, 533-543DOI: (10.1016/j.chempr.2018.01.005) Copyright © 2018 Terms and Conditions

Figure 3 Structure of the WT and Best Peptide (A) De novo structural reconstruction using PEP-FOLD3. (B) Structural characterization using CD spectroscopy. The spectra of the WT and best peptides are shown in yellow and red, respectively. (C) Helical wheel projection of peptides. Positively charged residues, hydrophobic residues, and other residues are shown in blue, white, and black, respectively. One-letter codes are used for amino acids. Chem 2018 4, 533-543DOI: (10.1016/j.chempr.2018.01.005) Copyright © 2018 Terms and Conditions

Scheme 1 General Schematic Describing the Evolutionary Process The circle represents the robotic process with the computational algorithm. In the first step, a random selection of the polymer sequence population is used as the starting polymer sequence population, and this forms the sequences experimentally synthesized in a peptide synthesis robot. The recorded polymer sequence activities then undergo analysis against a user-desired property (e.g., IC50) in the evaluation step. The sequences are ranked in terms of the desired property automatically, and those of poorest quality are rejected in the ranking step, allowing a new population to be selected. Meanwhile, the accepted sequences are used as a basis for creating a new sequence population after random mutation and crossover. Chem 2018 4, 533-543DOI: (10.1016/j.chempr.2018.01.005) Copyright © 2018 Terms and Conditions

Scheme 2 Searching for Potent Antimicrobial Peptides through a Workflow that Combines the Genetic Algorithm with Machine Learning Chem 2018 4, 533-543DOI: (10.1016/j.chempr.2018.01.005) Copyright © 2018 Terms and Conditions